v3 vs v4: choosing a computer use model on the API

What CUA versions mean

The cua_version field in POST /v1/runs sets which behavior you want. v3 is a controlled, step-by-step agent. You start a run and the API returns actions one at a time. You can pause, cancel, or resume. v4 is autonomous. The agent runs to completion without asking. It uses a built-in pass/fail verifier to decide if the task succeeded. Both versions still require a machine_id to drive a real desktop, browser, or terminal.

v3: step-by-step control

• ●Use cua_version = "v3" for tight control over each agent step.
• ●The agent returns actions incrementally. You can inspect or modify them before acting.
• ●You can pause a run with on_awaiting_human = "pause" and resume later.
• ●You can cancel or resume by ID with POST /v1/runs/{id}/cancel and POST /v1/runs/{id}/resume.
• ●Billed $0.05 per agent step with POST /v1/runs.

v4: autonomous mode

• ●Set cua_version = "v4" for an autonomous agent.
• ●The agent runs to completion without asking for human input.
• ●A pass/fail verifier decides if the task succeeded at the end.
• ●You can still set on_awaiting_human = "pause" or "fail" if you want to handle waiting steps explicitly.
• ●Billed $0.05 per agent step with POST /v1/runs.

Where this beats brittle automation

Traditional automation relies on fragile selectors and hard-coded APIs. Computer use agents see the screen and act like humans. They click buttons, fill forms, and navigate menus just as a person would. This works even when layouts change. It works when APIs are missing or undocumented. v3 lets you inspect and adjust actions. v4 lets you offload the whole flow to an autonomous agent that checks pass/fail. Both versions drive real desktops, browsers, and terminals, not just API calls.

Which version to choose

Choose v3 when you want to review each action or need to pause for human input. Choose v4 when you can define a clear success criterion and want the agent to run to completion. v4 can reduce development time for well-defined tasks. v3 is safer for exploratory work or when you need to intervene. Both versions bill $0.05 per agent step. Use the right version to balance control and autonomy.